Fix PhotulatorModelStack.fluxes; guard against legacy n_layers-1 artifacts

[ixkael]

What this fixes

1. PhotulatorModelStack.fluxes() called a nonexistent per-emulator method.

PhotulatorModelStack.fluxes() called self.emulators[i].fluxes(theta, N),
but neither Photulator nor PhotulatorBasic define a fluxes() method —
only the singular flux(theta, N) (returns flux in maggies). Any caller
using the stack's fluxes() path hit:

AttributeError: 'Photulator' object has no attribute 'fluxes'

One-line fix: call .flux() instead of .fluxes(). No tests/ layout
exists on this branch, so the commit message includes a doctest-style usage
note instead of a unit test.

2. Guard against loading legacy n_layers-1 Photulator artifacts.

Two incompatible generations of trained Photulator artifact exist:

• Current-generation artifacts were trained with the parametrized
  activation applied to every layer, including the output projection, and
  carry n_layers alpha/beta pairs — matching Photulator.forward()'s
  zip(self.W, self.b, self.alphas, self.betas) loop.
• Older, TF-converted artifacts trained a linear output layer and
  carry only n_layers - 1 alpha/beta pairs — no activation was ever
  trained (or saved) for the output layer.

Loading the latter through PhotulatorModelStack previously produced no
error. forward()'s zip loop silently truncates to the shortest input:
depending on how the artifact was constructed, this either drops the final
linear projection from the computation entirely, or pairs it with
uninitialized, randomly-drawn activation parameters that are never
overwritten. Either way, output magnitudes are wrong with no warning. On a
real (non-mock) artifact this produced ~48.6 mag of disagreement against
the correct value, with 2.3e-5 agreement recovered once the forward pass
for that artifact was redone with a linear (unactivated) final layer — the
artifact itself was fine, only the forward convention applied to it was
wrong.

This PR adds _check_photulator_layer_parameters(), called from
PhotulatorModelStack.__init__() right after each torch.load(). It
compares the alpha/beta arrays an artifact actually supplies (as loaded)
against its own n_layers, and raises a ValueError naming the file and
pointing to a linear-output forward pass (out = out @ W[-1] + b[-1], no
activation on the final layer) as the correct evaluation path for such
artifacts, instead of silently loading. PhotulatorBasic (no
alphas/betas; fixed, untrainable activation) is left untouched by the
check.

Testing

No tests/ directory exists on this branch. Manually verified:

• A toy Photulator, saved/loaded through PhotulatorModelStack, has
  stack.fluxes() succeed and match torch.cat([e.flux(theta, N) for e in stack.emulators]).
• A toy Photulator with alphas/betas truncated to n_layers - 1
  (simulating a legacy artifact) now raises ValueError at
  PhotulatorModelStack load time instead of silently loading.
• A normal n_layers-pair artifact is unaffected by the new check.

🤖 Generated with Claude Code